autodistill  by autodistill

Tool for training supervised models using foundation models, no labeling needed

created 2 years ago
2,354 stars

Top 19.9% on sourcepulse

GitHubView on GitHub
Project Summary

Autodistill enables users to train custom computer vision models without manual data labeling by leveraging large foundation models. It targets developers and researchers seeking to rapidly deploy efficient, specialized models for edge or cloud inference, bypassing the traditional bottleneck of data annotation.

How It Works

Autodistill employs a distillation pipeline: a large, capable "Base Model" (e.g., Grounding SAM, LLaVA) processes unlabeled images using an "Ontology" to generate auto-labeled datasets. These datasets then train a smaller, faster "Target Model" (e.g., YOLOv8, DETR), resulting in a deployable "Distilled Model." This approach democratizes model training by reducing reliance on human annotators and expensive labeling services.

Quick Start & Requirements

  • Install via pip: pip install autodistill autodistill-grounded-sam autodistill-yolov8
  • Requires Python 3.8+.
  • Example command: autodistill images --base="grounding_dino" --target="yolov8" --ontology '{"prompt": "label"}' --output="./dataset"
  • Colab Notebook: how-to-auto-train-yolov8-model-with-autodistill.ipynb

Highlighted Details

  • Supports object detection, instance segmentation, and classification tasks.
  • Extensive compatibility table lists numerous Base and Target models (e.g., Grounding DINO, SAM-CLIP, YOLOv8, DETR).
  • Pluggable interface allows easy integration of new models.
  • Optional deployment to Roboflow for edge and cloud applications.

Maintenance & Community

  • Actively developed by the Roboflow team.
  • Community resources include tutorials, guides, and a roadmap.
  • Open to contributions via a contributing guide.

Licensing & Compatibility

  • The core autodistill package is licensed under Apache 2.0.
  • Individual Base and Target model plugins may have their own licenses; users must check each plugin.
  • Generally compatible with commercial use, provided underlying model licenses permit.

Limitations & Caveats

  • Performance and accuracy depend heavily on the chosen Base Model and Ontology configuration.
  • Some model integrations are marked as "work in progress."
  • The project is positioned as an evolving system with ongoing development.
Health Check
Last commit

2 months ago

Responsiveness

1 week

Pull Requests (30d)
0
Issues (30d)
0
Star History
109 stars in the last 90 days

Explore Similar Projects

Feedback? Help us improve.